The Trade-offs of Domain Adaptation for Neural Language Models